crfm-helm 0.4.0__py3-none-any.whl → 0.5.10__py3-none-any.whl

helm/benchmark/metrics/basic_metrics.py

@@ -1,24 +1,19 @@
+from collections import defaultdict
 import math
-from dataclasses import dataclass, replace
-from typing import List, Callable, Optional, Dict, Tuple, Set, cast
+from dataclasses import dataclass
+from typing import List, Dict, Set
 from urllib.parse import unquote
-from functools import partial
 
-import json
-import string
-import nltk
 import numpy as np
-import re
-import scipy
-import calibration as cal
-import importlib_resources as resources
-from nltk.metrics.scores import f_measure
-from nltk.tokenize import word_tokenize
-from nltk.translate.bleu_score import sentence_bleu
-from rouge_score import rouge_scorer
+import scipy  # type: ignore
+import calibration as cal  # type: ignore
+from helm.benchmark.adaptation.scenario_state import ScenarioState
+from helm.benchmark.metrics.evaluate_reference_metrics import compute_reference_metrics, get_reference_metrics_metadata
+from helm.benchmark.metrics.efficiency_metrics import EfficiencyMetric
+from helm.benchmark.metrics.reference_metric import ReferenceMetric
 
 from helm.common.hierarchical_logger import hlog
-from helm.common.request import Token, Sequence
+from helm.common.request import Token, GeneratedOutput
 from helm.benchmark.adaptation.adapters.adapter_factory import (
     ADAPT_MULTIPLE_CHOICE_SEPARATE_ORIGINAL,
     ADAPT_MULTIPLE_CHOICE_SEPARATE_CALIBRATED,
@@ -29,196 +24,18 @@ from helm.benchmark.adaptation.adapter_spec import AdapterSpec
 from helm.benchmark.window_services.window_service import WindowService
 from helm.benchmark.window_services.window_service_factory import WindowServiceFactory
 from helm.benchmark.window_services.tokenizer_service import TokenizerService
-from helm.benchmark.scenarios.scenario import CORRECT_TAG, Instance, Reference
-from helm.benchmark.scenarios.math_scenario import is_equiv, is_equiv_chain_of_thought
-from helm.benchmark.scenarios.code_scenario import CodeReference
-from helm.benchmark.metrics.cleva_metrics_helper import ChineseTokenizer
-from . import code_metrics_helper
-from .metric import Metric, get_unique_stat_by_name
-from .metric_name import MetricName
-from .metric_service import MetricService
-from .statistic import Stat
-
-
-try:
-    nltk.data.find("tokenizers/punkt")
-except LookupError:
-    nltk.download("punkt")  # Required for rouge
-
-
-EFFICIENCY_DATA_PACKAGE: str = "helm.benchmark.efficiency_data"
-
-INFERENCE_IDEALIZED_RUNTIMES_JSON_FILENAME: str = "inference_idealized_runtimes.json"
-INFERENCE_DENOISED_RUNTIMES_JSON_FILENAME: str = "inference_denoised_runtimes.json"
-TRAINING_EFFICIENCY_JSON_FILENAME: str = "training_efficiency.json"
-
-
-def compute_estimated_time_from_prompt_size_and_num_output_tokens(
-    request_state: RequestState,
-    inference_runtimes_dict: Dict[str, Dict],
-    num_prompt_tokens: int,
-    num_output_tokens: int,
-) -> Optional[float]:
-    estimated_runtime: Optional[float]
-    if request_state.request.model_deployment in inference_runtimes_dict:
-        inference_runtimes_dict_for_model = inference_runtimes_dict[request_state.request.model_deployment]
-        runtime_per_output_token: float = inference_runtimes_dict_for_model["runtime_per_output_token"]
-        raw_runtimes_for_prompt_tokens: Dict[str, float] = inference_runtimes_dict_for_model[
-            "runtime_for_prompt_tokens"
-        ]
-        runtimes_for_prompt_tokens: Dict[int, float] = {int(k): v for (k, v) in raw_runtimes_for_prompt_tokens.items()}
-
-        runtime_for_prompt_tokens: Optional[float] = None
-        largest_num_tokens_in_efficiency_dict: int = max(runtimes_for_prompt_tokens.keys())
-        # Find the smallest num_prompt_tokens larger than the number of tokens in the given prompt,
-        # then scale runtime in dict by (num_prompt_tokens / key) to get more accurate estimate: we
-        # assume that we can encode the prompt at the same throughput as the smallest key larger than
-        # num_prompt_tokens, and number of compute operations scales linearly with num_prompt_tokens.
-        for key in sorted(runtimes_for_prompt_tokens.keys()):
-            if num_prompt_tokens <= key:
-                runtime_for_prompt_tokens = runtimes_for_prompt_tokens[key] * (num_prompt_tokens / key)
-                break
-        # If number of tokens in the prompt exceeds the largest key in the efficiency dict, then
-        # estimate the prompt encoding time by linearly scaling up the runtime for the largest
-        # key (this is reasonably accurate under certain simplifying assumptions).
-        if runtime_for_prompt_tokens is None:
-            runtime_for_prompt_tokens = runtimes_for_prompt_tokens[largest_num_tokens_in_efficiency_dict] * (
-                num_prompt_tokens / largest_num_tokens_in_efficiency_dict
-            )
-        overhead: Optional[float] = inference_runtimes_dict_for_model.get("overhead")
-
-        # Idealized runtime is sum of the runtime of encoding the input tokens, the runtime of
-        # generating `num_output_tokens` (`runtime_per_output_token` * (`num_output_tokens` - 1))
-        # if number of output tokens is greater than 0, otherwise just `runtime_for_prompt_tokens`,
-        # and the overhead if available.
-        estimated_runtime = runtime_for_prompt_tokens
-        if num_output_tokens > 0:
-            estimated_runtime += runtime_per_output_token * (num_output_tokens - 1)
-        # Add overhead if it is available.
-        if overhead is not None:
-            estimated_runtime += overhead
-    else:
-        estimated_runtime = None
-
-    return estimated_runtime
-
-
-def pass_at_k_estimator(n: int, c: int, k: int) -> float:
-    """Calculates 1 - comb(n - c, k) / comb(n, k).
-
-    Numerically stable version defined in
-        https://arxiv.org/pdf/2107.03374.pdf
-    """
-    if n - c < k:
-        return 1.0
-    return 1.0 - np.prod(1.0 - k / np.arange(n - c + 1, n + 1))
-
-
-def normalize_text(text: str) -> str:
-    """Lower text and remove punctuation, articles and extra whitespace.
-    Copied from the [QuAC](http://quac.ai/) evaluation script found at
-    https://s3.amazonaws.com/my89public/quac/scorer.py"""
-
-    def remove_articles(text: str) -> str:
-        return re.sub(r"\b(a|an|the)\b", " ", text)
-
-    def white_space_fix(text: str) -> str:
-        return " ".join(text.split())
-
-    def remove_punc(text: str) -> str:
-        exclude = set(string.punctuation)
-        return "".join(ch for ch in text if ch not in exclude)
-
-    def lower(text: str) -> str:
-        return text.lower()
-
-    return white_space_fix(remove_articles(remove_punc(lower(text))))
-
-
-def exact_match(gold: str, pred: str) -> float:
-    if not pred:
-        return 0
-
-    return 1 if gold.strip() == pred.strip() else 0
-
-
-def quasi_exact_match(gold: str, pred: str) -> float:
-    if not pred:
-        return 0
-
-    return 1 if normalize_text(gold) == normalize_text(pred) else 0
-
-
-def prefix_exact_match(gold: str, pred: str) -> float:
-    """
-    The `prefix_exact_match` metric is particularly useful in the zero-shot setting, where the model is
-    not given examples of the expected outputs and tends to output more tokens than it should.
-
-    For example, for this zero-shot prompt from BoolQ,
-
-    Passage: Elmendorf Air Force Base (IATA: EDF, ICAO: PAED, FAA LID: EDF) is a United States military facility
-    in Anchorage, the largest city in Alaska. Originally known as Elmendorf Field, it became Elmendorf Air Force
-    Base after World War II, and in 2010 it merged with nearby Fort Richardson to form Joint Base Elmendorf-Richardson.
-    Question: Is there an air force base in anchorage alaska?
-    Answer:
-
-    the model could output up to `max_tokens` number of tokens "Yes, Elmendorf" instead of just "Yes".
-    """
-    if not pred:
-        return 0
-
-    return 1 if pred.strip().startswith(gold.strip()) else 0
-
-
-def quasi_prefix_exact_match(gold: str, pred: str) -> float:
-    """
-    Same thing as `prefix_exact_match` but we normalize the text before checking if the prefix match.
-    """
-    if not pred:
-        return 0
-
-    return 1 if normalize_text(pred).startswith(normalize_text(gold)) else 0
-
-
-def f1_score(gold: str, pred: str) -> float:
-    ret = f_measure(set(normalize_text(gold).split()), set(normalize_text(pred).split()))
-    if ret is None:  # answer is the empty string after normalizing
-        return 0.0
-
-    return ret
-
-
-def exact_match_indicator(gold: str, pred: str, indicator: str = " ") -> float:
-    """
-    Exact match, allowing for some preceding context.
-    For example, the following two answers are considered matching:
-    - Because of x and y, the answer is ## <answer>
-    - Given reasons y and z, the answer is ## <answer>
-    While the following is considered different from the earlier two
-    - Given reasons x and a, the answer is ## <other answer>
-    """
-    pred = pred.split(indicator)[-1].strip()
-    gold = gold.split(indicator)[-1].strip()
-    return exact_match(gold, pred)
-
-
-def final_number_exact_match(gold: str, pred: str) -> float:
-    """
-    Returns 1 iff the final number in gold and pred match.
-    Similar to exact_match_indicator.
-    Example:
-    - gold = "The answer is 15."
-    - pred = "The answer is 15 eggs."
-    - Returns 1
-    """
-
-    def get_final_number(x: str) -> str:
-        matches = re.findall(r"-?[\d,]+(?:.\d+)?", x)
-        if not matches:
-            return ""
-        return matches[-1].replace(",", "")
-
-    return exact_match(get_final_number(gold), get_final_number(pred))
+from helm.benchmark.scenarios.scenario import CORRECT_TAG, Instance
+from helm.benchmark.metrics.metric import (
+    Metric,
+    MetricInterface,
+    MetricMetadata,
+    MetricResult,
+    add_context,
+    get_unique_stat_by_name,
+)
+from helm.benchmark.metrics.metric_name import MetricContext, MetricName
+from helm.benchmark.metrics.metric_service import MetricService
+from helm.benchmark.metrics.statistic import Stat, merge_stat
 
 
 def get_num_bytes(tokens: List[Token]) -> int:
@@ -270,123 +87,6 @@ def convert_tokens_to_text(tokens: List[Token]) -> List[Dict]:
     return groups
 
 
-def rouge_score(gold: str, pred: str, rouge_type: str, scorer: rouge_scorer.RougeScorer) -> float:
-    scores = scorer.score(gold, pred)
-    return scores[rouge_type].fmeasure
-
-
-def get_rouge_function(rouge_type: str) -> Callable[[str, str], float]:
-    scorer = rouge_scorer.RougeScorer([rouge_type], use_stemmer=True)
-    return partial(rouge_score, scorer=scorer, rouge_type=rouge_type)
-
-
-def bleu_1(gold: str, pred: str) -> float:
-    return sentence_bleu([word_tokenize(gold)], word_tokenize(pred), weights=(1, 0, 0, 0))
-
-
-def chinese_bleu_1(gold: str, pred: str) -> float:
-    char_tokenizer = ChineseTokenizer()
-    return sentence_bleu([char_tokenizer.tokenize(gold)], char_tokenizer.tokenize(pred), weights=(1, 0, 0, 0))
-
-
-def get_chinese_rouge_function(rouge_type: str) -> Callable[[str, str], float]:
-    char_tokenizer = ChineseTokenizer()
-    scorer = rouge_scorer.RougeScorer([rouge_type], use_stemmer=True, tokenizer=char_tokenizer)
-    return partial(rouge_score, scorer=scorer, rouge_type=rouge_type)
-
-
-def cleva_math_result_match(gold: str, pred: str) -> float:
-    """
-    Exact match that only cares the last math expression.
-    Common math expressions are numbers and fractions.
-    """
-    pattern = r"[-+*/%\.\(\)\d]+"
-    matches = re.findall(pattern, pred)
-    if matches:
-        pred = matches[-1].lstrip(")")
-    # remove space in front or at the end
-    pred = pred.strip()
-    return exact_match(gold, pred)
-
-
-def bleu_4(gold: str, pred: str) -> float:
-    return sentence_bleu([word_tokenize(gold)], word_tokenize(pred), weights=(0, 0, 0, 1))
-
-
-def extract_set_from_text(
-    set_str: str,
-    set_start_str: str = " is ",
-    set_separator: str = " and ",
-    empty_set_str: str = "Nothing.",
-) -> Set[str]:
-    """
-    Given a string, extract the set of strings implied by that string.
-    set_start_str denotes the start of the set
-    set_separator denotes the string separating set elements
-    empty_set_str is the string which denotes the empty set
-    """
-    if set_str == empty_set_str:
-        return set()
-    set_str = set_str.replace(".", "")
-    extracted_set = set(set_str.split(set_start_str)[-1].split(set_separator))
-    return extracted_set
-
-
-def extract_gold_pred_sets(gold: str, pred: str) -> Tuple[Set[str], Set[str]]:
-    """Extract the set of strings implied by the gold and pred strings"""
-    gold_set = extract_set_from_text(gold)
-    pred_set = extract_set_from_text(pred.split("\n")[0])
-    return gold_set, pred_set
-
-
-def iou_set_match(gold: str, pred: str) -> float:
-    """Compute the intersection over union of the gold and pred sets"""
-    gold_set, pred_set = extract_gold_pred_sets(gold, pred)
-    if len(gold_set) == 0:  # If gold is empty, just check if the pred set is also empty
-        return float(gold_set == pred_set)
-    return len(gold_set.intersection(pred_set)) / len(gold_set.union(pred_set))
-
-
-def f1_set_match(gold: str, pred: str) -> float:
-    """Compute the F1 score of the gold and pred sets"""
-    gold_set, pred_set = extract_gold_pred_sets(gold, pred)
-    if len(gold_set) == 0:  # If gold is empty, just check if the pred set is also empty
-        return float(gold_set == pred_set)
-    true_positives = gold_set.intersection(pred_set)
-    return 2 * len(true_positives) / (len(gold_set) + len(pred_set))
-
-
-def exact_set_match(gold: str, pred: str) -> float:
-    """Compute whether the sets generated exactly match"""
-    gold_set, pred_set = extract_gold_pred_sets(gold, pred)
-    return float(gold_set == pred_set)
-
-
-def absolute_value_difference(gold: str, pred: str) -> float:
-    """Compute the absolute value of the difference between two numbers (provided as strings),
-    or 0.0 if invalid input.
-    """
-
-    def maybe_int(text: str):
-        """Parse int, ignoring commas in numbers."""
-        try:
-            val = int(text.replace(",", ""))
-        except ValueError:
-            return 0.0
-        return val
-
-    gold_val = maybe_int(gold)
-    pred_val = maybe_int(pred)
-    return abs(gold_val - pred_val)
-
-
-def code_eval(gold: Tuple[str, Optional[Dict]], pred: str) -> float:
-    """Evaluate Code Correctness on test examples."""
-    assert gold[1] is not None  # gold[1]["canonical_solution"]
-    # Warning: will execute machine generated code; need to sandbox before executing
-    return float(code_metrics_helper.check_correctness(gold[1], pred, 3.0)["passed"])  # type: ignore
-
-
 def compute_perplexity_metrics(stats: Dict[MetricName, Stat]) -> List[Stat]:
     # TODO: find out the root cause and undo num_X > 0 check
     #       https://github.com/stanford-crfm/benchmarking/issues/350
@@ -411,7 +111,76 @@ def compute_perplexity_metrics(stats: Dict[MetricName, Stat]) -> List[Stat]:
     return derived_stats
 
 
-class BasicMetric(Metric):
+def _get_perplexity_metrics_metadata() -> List[MetricMetadata]:
+    return [
+        MetricMetadata(
+            name="perplexity",
+            display_name="Perplexity",
+            short_display_name="PPL",
+            description="Perplexity of the output completion (effective branching factor per output token).",
+            lower_is_better=True,
+            group=None,
+        ),
+        MetricMetadata(
+            name="logprob_per_byte",
+            display_name="Log probability / byte",
+            short_display_name="Logprob/byte",
+            description="Predicted output's average log probability normalized by the number of bytes.",
+            lower_is_better=False,
+            group=None,
+        ),
+        MetricMetadata(
+            name="bits_per_byte",
+            display_name="Bits/byte",
+            short_display_name="BPB",
+            description="Average number of bits per byte according to model probabilities.",
+            lower_is_better=True,
+            group=None,
+        ),
+    ]
+
+
+class InstancesPerSplitMetric(MetricInterface):
+    """Report the average num_instances in each MetricContext across train_trials."""
+
+    def evaluate(
+        self, scenario_state: ScenarioState, metric_service: MetricService, eval_cache_path: str, parallelism: int
+    ) -> MetricResult:
+        adapter_spec = scenario_state.adapter_spec
+        global_stats: Dict[MetricName, Stat] = {}
+
+        for train_trial_index in range(adapter_spec.num_train_trials):
+            trial_stats: Dict[MetricName, Stat] = {}  # Statistics just for this trial
+            # Group instances in this train_trial by context.
+            instances_per_metric_context: Dict[MetricContext, Set[Instance]] = defaultdict(set)
+            for request_state in scenario_state.request_states:
+                if request_state.train_trial_index == train_trial_index:
+                    instances_per_metric_context[MetricContext.from_instance(request_state.instance)].add(
+                        request_state.instance
+                    )
+            for context, instance_set in instances_per_metric_context.items():
+                stat = Stat(MetricName("num_instances")).add(len(instance_set))
+                merge_stat(trial_stats, add_context(stat, context))
+
+            # We take the mean value for each trial.
+            for stat in trial_stats.values():
+                merge_stat(global_stats, stat.take_mean())
+
+        # There are no per-instance Stats.
+        return MetricResult(list(global_stats.values()), [])
+
+    def get_metadata(self) -> List[MetricMetadata]:
+        return [
+            MetricMetadata(
+                name="num_instances",
+                display_name="# eval",
+                description="Number of evaluation instances.",
+                lower_is_better=None,
+            )
+        ]
+
+
+class BasicGenerationMetric(Metric):
     """
     Defines basic metrics which don't require domain knowledge.  This should be
     fairly comprehensive already, and we should try to use this as much as possible.
@@ -422,338 +191,11 @@ class BasicMetric(Metric):
 
     def __init__(self, names: List[str]):
         self.names: List[str] = names
-
-        # For Efficiency metrics:
-        # The `inference_efficiency.json` file contains a `runtime_per_output_token` value
-        # (the estimated runtime of generating one output token) and a
-        # `runtime_for_prompt_tokens` dict (a mapping from various num_prompt_tokens values to
-        # the estimated runtime of encoding a prompt with that many tokens).
-        # For example:
-        # "openai/davinci": {
-        #   "runtime_per_output_token": 0.080,
-        #   "runtime_for_prompt_tokens": {
-        #     "1": 0.016,
-        #     "16": 0.018,
-        #     "32": 0.020,
-        #     ...
-        #
-        # These runtimes are generated by initializing Megatron with a model of the right size,
-        # obtaining end-to-end generation times for different numbers of prompt and output tokens,
-        # and then fitting a linear regression model to the runtimes: the resulting slope is the
-        # runtime_per_output_token, which is the processing time for generating each output token,
-        # and the y-intercept is the runtime_for_prompt_tokens, with different values for different
-        # num_prompt_tokens values.
-        # Profiling code and logs, and code to fit the regression model is available at
-        # https://github.com/stanford-crfm/benchmarking_efficiency.
-        data_package = resources.files(EFFICIENCY_DATA_PACKAGE)
-        with data_package.joinpath(INFERENCE_IDEALIZED_RUNTIMES_JSON_FILENAME).open("r") as f:
-            self.inference_idealized_runtimes_dict = json.load(f)
-        with data_package.joinpath(INFERENCE_DENOISED_RUNTIMES_JSON_FILENAME).open("r") as f:
-            self.inference_denoised_runtimes_dict = json.load(f)
-
-        # We use estimated emitted CO2 during training (in tons of CO2) as a proxy metric
-        # for training efficiency. We use reported metrics where applicable, otherwise
-        # we estimate them from runtime information, type and number of hardware accelerators
-        # used, region, etc.
-        with data_package.joinpath(TRAINING_EFFICIENCY_JSON_FILENAME).open("r") as f:
-            self.training_efficiency_dict = json.load(f)
+        self.efficiency_metric = EfficiencyMetric()
 
     def __repr__(self):
         return f"BasicMetric({','.join(self.names)})"
 
-    def compute_reference_metrics(
-        self, adapter_spec: AdapterSpec, request_state: RequestState, metric_service: MetricService
-    ) -> List[Stat]:
-        """
-        Setup:
-
-        - Gold (correct references): G1 ... Gm
-        - Predictions (completions): P1 ... Pk
-
-        For each pair (G, P), we can define a ${score} (e.g., exact match, F1, BLEU).
-
-        We define the following stats:
-
-        - ${score}: max_i score(Gi, P1)
-        - ${score}@k: max_{i,j} score(Gi, Pj)
-        """
-
-        def compute_metrics_helper(
-            name: MetricName,
-            score_func: Callable,
-            group: Optional[str] = None,
-        ) -> List[Stat]:
-            if name.name == "pass":  # Calculate pass@k for HumanEval from CodeScenario.
-                score_func = cast(Callable[[Tuple[str, Optional[Dict]], str], float], score_func)  # Make mypy happy.
-                code_golds = cast(List[CodeReference], golds)
-                results = [
-                    score_func((gold.output.text, gold.test_cases), pred) for gold in code_golds for pred in preds
-                ]
-                _len, _sum = len(results), int(sum(results))  # Cast to int to make type match.
-                score_1 = pass_at_k_estimator(_len, _sum, 1)
-                score_k = pass_at_k_estimator(_len, _sum, adapter_spec.num_outputs)
-            elif name.name == "code_eval_acc":
-                score_func = cast(Callable[[Tuple[str, Optional[Dict]], str], float], score_func)  # Make mypy happy.
-                code_golds = cast(List[CodeReference], golds)
-                score_1 = max(score_func((gold.output.text, gold.test_cases), preds[0]) for gold in code_golds)
-                score_k = max(
-                    score_func((gold.output.text, gold.test_cases), pred) for gold in code_golds for pred in preds
-                )
-            else:
-                score_func = cast(Callable[[str, str], float], score_func)  # Make mypy happy.
-                score_1 = max(score_func(gold.output.text, preds[0]) for gold in golds)
-                score_k = max(score_func(gold.output.text, pred) for gold in golds for pred in preds)
-
-            metrics = [Stat(name).add(score_1)]  # score_1 corresponds using one prediction
-            if adapter_spec.num_outputs != 1:
-                metrics.append(Stat(replace(name, name=f"{name.name}@{adapter_spec.num_outputs}")).add(score_k))
-            return metrics
-
-        # maps each string metric name to its associated function
-        metric_fn_mapping: Dict[str, Callable] = {
-            "exact_match": exact_match,
-            "quasi_exact_match": quasi_exact_match,
-            "prefix_exact_match": prefix_exact_match,
-            "quasi_prefix_exact_match": quasi_prefix_exact_match,
-            "exact_match_indicator": exact_match_indicator,
-            "final_number_exact_match": final_number_exact_match,
-            "exact_set_match": exact_set_match,
-            "iou_set_match": iou_set_match,
-            "f1_set_match": f1_set_match,
-            "math_equiv": is_equiv,
-            "math_equiv_chain_of_thought": is_equiv_chain_of_thought,
-            "code_eval_acc": code_eval,
-            "pass": code_eval,
-            "f1_score": f1_score,
-            "rouge_1": get_rouge_function("rouge1"),
-            "rouge_2": get_rouge_function("rouge2"),
-            "rouge_l": get_rouge_function("rougeL"),
-            "bleu_1": bleu_1,
-            "bleu_4": bleu_4,
-            "chinese_bleu_1": chinese_bleu_1,
-            "chinese_rouge_1": get_chinese_rouge_function("rouge1"),
-            "chinese_rouge_2": get_chinese_rouge_function("rouge2"),
-            "cleva_math_result_match": cleva_math_result_match,
-            "absolute_value_difference": absolute_value_difference,
-        }
-
-        stats: List[Stat] = []
-
-        # Gold outputs
-        golds: List[Reference] = [reference for reference in request_state.instance.references if reference.is_correct]
-        assert len(golds) > 0
-
-        # Predicted outputs
-        assert request_state.result is not None
-        sorted_completions: List[Sequence] = sorted(request_state.result.completions, key=lambda x: -x.logprob)
-        preds: List[str] = [completion.text.strip() for completion in sorted_completions]
-
-        # Apply mapping if exists (e.g., for multiple-choice questions A -> Boston, B -> New York)
-        # Note: If 'A' and 'B' were the only possible choices, smaller language models like GPT-2 would
-        # sometimes predict a random letter like 'M'.
-        if request_state.output_mapping is not None:
-            preds = [request_state.output_mapping.get(pred) for pred in preds]  # type: ignore
-
-        # Compute max_prob, the probability that the model assigns to its generated text.
-        # Use the log prob of sorted_completions[0], which is the completion with the highest
-        # log_prob. We use this since that's what's used for computing metrics like exact_match.
-        # One subtlety is that when computing exact_match, we strip whitespace, so the actual
-        # max_prob is the sum of all the probabilities in the set {x : strip(x) = prediction}.
-        # In practice, we think this may not make much of a difference because models may not place
-        # high probabilities on having additional spaces (should check this). Also, the sum
-        # involves computing the log_prob for many completions which could be intractable.
-        max_prob = np.exp(sorted_completions[0].logprob)
-        stats.append(Stat(MetricName("max_prob")).add(max_prob))
-
-        # Add other metrics
-        for metric_name in self.names:
-            if metric_name in metric_fn_mapping:
-                stats.extend(compute_metrics_helper(MetricName(metric_name), metric_fn_mapping[metric_name]))
-            else:
-                raise NameError(f"{metric_name} is not in the list of metric functions.")
-
-        return stats
-
-    def compute_efficiency_metrics(
-        self, adapter_spec: AdapterSpec, request_state: RequestState, metric_service: MetricService
-    ) -> List[Stat]:
-        """Compute efficiency metrics for both inference and training.
-        For inference, we record both the actual runtime and an estimated idealized runtime
-        for the given request with an optimized software implementation run on A100 GPU(s),
-        taking into account both the number of tokens in the prompt of the request, and the
-        number of generated output tokens.
-        For training, we report the estimated total metric tons of CO2 emitted to train the
-        model. This is the same for each request."""
-        # Compute efficiency metrics for inference.
-        assert request_state.result is not None
-
-        runtime: Optional[float] = None
-        batch_size: Optional[int] = None
-        # Compute efficiency metrics for inference.
-        if request_state.result.request_time is not None:
-            runtime = request_state.result.request_time
-            batch_size = 1
-        # For models that perform offline batch inference, effective runtime is batch_request_time, but also
-        # record batch_size to provide nuance.
-        if request_state.result.batch_request_time is not None and request_state.result.batch_size is not None:
-            runtime = request_state.result.batch_request_time
-            batch_size = request_state.result.batch_size
-
-        # Compute total number of prompt and output tokens.
-        # Fetch the right `Tokenizer` depending on the model defined in `AdapterSpec`
-        # and calculate the number of tokens in the prompt.
-        tokenizer_service: TokenizerService = metric_service
-        window_service: WindowService = WindowServiceFactory.get_window_service(
-            adapter_spec.model_deployment, tokenizer_service
-        )
-        prompt: str = request_state.request.prompt
-        num_prompt_tokens: int = window_service.get_num_tokens(prompt)
-
-        # Total number of tokens in the completion.
-        num_completion_tokens: int = sum([len(completion.tokens) for completion in request_state.result.completions])
-        # Don't include prompt in number of generated tokens (e.g., for language modeling).
-        # Assume that tokens for different completions are generated sequentially (instead of batched) when
-        # computing num_output_tokens (for the purpose of runtime estimation).
-        num_output_tokens: int = num_completion_tokens
-        if request_state.request.echo_prompt:
-            # num_prompt_tokens > num_output_tokens can happen if tokenizer doesn't round trip.
-            if num_prompt_tokens <= num_output_tokens:
-                num_output_tokens -= num_prompt_tokens
-            else:
-                hlog(
-                    f"WARNING: num_prompt_tokens ({num_prompt_tokens}) > num_output_tokens ({num_output_tokens}) "
-                    f"for prompt: {prompt}"
-                )
-                num_output_tokens = 0
-
-        idealized_runtime: Optional[float] = compute_estimated_time_from_prompt_size_and_num_output_tokens(
-            request_state, self.inference_idealized_runtimes_dict, num_prompt_tokens, num_output_tokens
-        )
-
-        denoised_runtime: Optional[float] = compute_estimated_time_from_prompt_size_and_num_output_tokens(
-            request_state, self.inference_denoised_runtimes_dict, num_prompt_tokens, num_output_tokens
-        )
-        # Denoised runtime for offline models is just runtime.
-        # We divide by batch_size to get approximate per-input runtime.
-        if runtime is not None and request_state.result.batch_size is not None:
-            denoised_runtime = runtime / request_state.result.batch_size
-
-        # Compute efficiency metrics for training.
-        training_co2_cost: Optional[float]
-        if request_state.request.model_deployment in self.training_efficiency_dict["carbon"]:
-            training_co2_cost = self.training_efficiency_dict["carbon"][request_state.request.model_deployment]["value"]
-        else:
-            training_co2_cost = None
-
-        training_energy_cost: Optional[float]
-        if request_state.request.model_deployment in self.training_efficiency_dict["energy"]:
-            training_energy_cost = self.training_efficiency_dict["energy"][request_state.request.model_deployment][
-                "value"
-            ]
-        else:
-            training_energy_cost = None
-
-        stats = [
-            Stat(MetricName("num_prompt_tokens")).add(num_prompt_tokens),
-            Stat(MetricName("num_completion_tokens")).add(num_completion_tokens),
-            Stat(MetricName("num_output_tokens")).add(num_output_tokens),
-            Stat(MetricName("training_co2_cost")).add(training_co2_cost),
-            Stat(MetricName("training_energy_cost")).add(training_energy_cost),
-        ]
-        if runtime is not None:
-            stats.append(Stat(MetricName("inference_runtime")).add(runtime))
-        if batch_size is not None:
-            stats.append(Stat(MetricName("batch_size")).add(batch_size))
-        if denoised_runtime is not None:
-            stats.append(Stat(MetricName("inference_denoised_runtime")).add(denoised_runtime))
-        if idealized_runtime is not None:
-            stats.append(Stat(MetricName("inference_idealized_runtime")).add(idealized_runtime))
-        return stats
-
-    def compute_finish_reason_metrics(
-        self, adapter_spec: AdapterSpec, request_state: RequestState, metric_service: MetricService
-    ) -> List[Stat]:
-        """Record how often generation finished due to reaching token limit, stop token(s), or end of text"""
-        assert request_state.result is not None
-        sequence = request_state.result.completions[0]
-        valid_reasons = [
-            "length",
-            "stop",
-            "endoftext",
-            "unknown",
-        ]
-        if sequence.finish_reason is None or sequence.finish_reason["reason"] not in valid_reasons:
-            reason = "unknown"
-        else:
-            reason = sequence.finish_reason["reason"]
-        return [
-            Stat(MetricName(f"finish_reason_{valid_reason}")).add(int(reason == valid_reason))
-            for valid_reason in valid_reasons
-        ]
-
-    def compute_truncation_metrics(
-        self, adapter_spec: AdapterSpec, request_state: RequestState, metric_service: MetricService
-    ) -> List[Stat]:
-        """
-        Record the number of training instances used in the prompt and whether
-        even the prompt needed to be truncated (once we hit zero training instances).
-        """
-        return [
-            Stat(MetricName("num_train_instances")).add(request_state.num_train_instances),
-            Stat(MetricName("prompt_truncated")).add(request_state.prompt_truncated),
-        ]
-
-    def compute_all_general_metrics(
-        self, adapter_spec: AdapterSpec, request_state: RequestState, metric_service: MetricService
-    ) -> List[Stat]:
-        """
-        Compute metrics that are common to both `evaluate_generation` and `evaluate_references`.
-        """
-        stats: List[Stat] = []
-
-        stats.append(Stat(MetricName("num_references")).add(len(request_state.instance.references)))
-
-        # Copy from adapter spec
-        stats.append(Stat(MetricName("num_train_trials")).add(adapter_spec.num_train_trials))
-
-        stats.extend(self.compute_efficiency_metrics(adapter_spec, request_state, metric_service))
-        stats.extend(self.compute_finish_reason_metrics(adapter_spec, request_state, metric_service))
-        stats.extend(self.compute_truncation_metrics(adapter_spec, request_state, metric_service))
-
-        return stats
-
-    def compute_language_modeling_metrics(
-        self, adapter_spec: AdapterSpec, request_state: RequestState, metric_service: MetricService
-    ) -> List[Stat]:
-        """Compute the logprob and normalization factors for the first completion"""
-        assert request_state.result is not None
-        sequence = request_state.result.completions[0]
-
-        # Remove the empty tokens (typically generated by the AI21 tokenizer in the beginning of the text)
-        #
-        # Some more details about AI21 tokenizer: If the input prompt begins with a space, then
-        # the tokenizer inserts an empty token to the beginning.
-        # e.g. " burying him" -> ["▁"(0,0), "▁burying"(0,8), "▁him"(8,12)].
-        # TODO(#1522): Update this comment once solved.
-        # Since this empty token is introduced by our chunking approach, we need to remove it.
-        tokens: List[Token]
-        if request_state.num_conditioning_tokens > 0 and sequence.tokens[0].text == "":
-            tokens = sequence.tokens[1:]
-        else:
-            tokens = sequence.tokens
-        pred_tokens = tokens[request_state.num_conditioning_tokens :]
-        logprob, num_perplexity_tokens, num_bytes = (
-            sum(token.logprob for token in pred_tokens),
-            len(pred_tokens),
-            get_num_bytes(pred_tokens),
-        )
-
-        return [
-            Stat(MetricName("logprob")).add(logprob),
-            Stat(MetricName("num_perplexity_tokens")).add(num_perplexity_tokens),
-            Stat(MetricName("num_bytes")).add(num_bytes),
-        ]
-
     def evaluate_generation(
         self,
         adapter_spec: AdapterSpec,
@@ -763,15 +205,49 @@ class BasicMetric(Metric):
     ) -> List[Stat]:
         """Compute all metrics."""
         stats: List[Stat] = []
-        stats.extend(self.compute_all_general_metrics(adapter_spec, request_state, metric_service))
+        stats.extend(compute_request_state_metrics(self.efficiency_metric, adapter_spec, request_state, metric_service))
 
         if len(request_state.instance.references) > 0:
-            stats.extend(self.compute_reference_metrics(adapter_spec, request_state, metric_service))
+            stats.extend(compute_reference_metrics(self.names, adapter_spec, request_state, metric_service))
 
-        stats.extend(self.compute_language_modeling_metrics(adapter_spec, request_state, metric_service))
+        stats.extend(compute_language_modeling_metrics(adapter_spec, request_state, metric_service))
 
         return stats
 
+    def derive_stats(self, stats_dict: Dict[MetricName, Stat]) -> List[Stat]:
+        """Derive perplexity metrics if applicable. We don't worry about splits and perturbations here."""
+        derived_stats: List[Stat] = []
+        derived_stats.extend(compute_perplexity_metrics(stats_dict))
+        return derived_stats
+
+    def derive_per_instance_stats(self, per_instance_stats: Dict[Instance, List[Stat]]) -> List[Stat]:
+        """Derive calibration metrics if applicable. We don't worry about splits and perturbations here."""
+        derived_stats: List[Stat] = []
+        derived_stats.extend(compute_calibration_metrics(per_instance_stats))
+        return derived_stats
+
+    def get_metadata(self) -> List[MetricMetadata]:
+        return (
+            get_request_state_metrics_metadata(self.efficiency_metric)
+            + get_reference_metrics_metadata(self.names)
+            + _get_language_modeling_metrics_metadata()
+            + _get_perplexity_metrics_metadata()
+            + _get_calibration_metrics_metadata()
+        )
+
+
+class BasicReferenceMetric(ReferenceMetric):
+    """
+    Defines basic metrics for Scenarios that use one Request per Reference instead of
+    one per Instance.
+    """
+
+    def __init__(self):
+        self.efficiency_metric = EfficiencyMetric()
+
+    def __repr__(self):
+        return "BasicReferenceMetric"
+
     def evaluate_references(
         self,
         adapter_spec: AdapterSpec,
@@ -801,7 +277,7 @@ class BasicMetric(Metric):
             assert len(request_state.result.completions) == 1
 
             reference_index = request_state.reference_index
-            sequence: Sequence = request_state.result.completions[0]
+            sequence: GeneratedOutput = request_state.result.completions[0]
             reference: str = request_state.instance.references[reference_index].output.text
 
             # Find the span of the completion that matches the reference.
@@ -848,8 +324,14 @@ class BasicMetric(Metric):
             raise ValueError(f"Unknown adapter method: {adapter_spec.method}")
 
         stats: List[Stat] = []
-        stats.extend(self.compute_all_general_metrics(adapter_spec, request_state, metric_service))
 
+        general_metrics: Dict[MetricName, Stat] = {}
+        for request_state in reference_request_states:
+            for stat in compute_request_state_metrics(
+                self.efficiency_metric, adapter_spec, request_state, metric_service
+            ):
+                merge_stat(general_metrics, stat)
+        stats.extend(general_metrics.values())
         max_prob = np.max(scipy.special.softmax(reference_scores))
 
         # Multiple references may attain the same maximal score; in such cases,
@@ -868,18 +350,229 @@ class BasicMetric(Metric):
         )
         return stats
 
-    def derive_stats(self, stats_dict: Dict[MetricName, Stat]) -> List[Stat]:
-        """Derive perplexity metrics if applicable. We don't worry about splits and perturbations here."""
-        derived_stats: List[Stat] = []
-        derived_stats.extend(compute_perplexity_metrics(stats_dict))
-        return derived_stats
+    def get_metadata(self) -> List[MetricMetadata]:
+        return [
+            MetricMetadata(
+                name="max_prob",
+                display_name="Max prob",
+                description="Model's average confidence in its prediction (only computed for classification tasks)",
+                lower_is_better=False,
+                group="calibration_detailed",
+            ),
+            MetricMetadata(
+                name="exact_match",
+                display_name="Exact match",
+                short_display_name="EM",
+                description="Fraction of instances that the predicted output matches a correct reference exactly.",
+                lower_is_better=False,
+                group="accuracy",
+            ),
+            MetricMetadata(
+                name="predicted_index",
+                display_name="Predicted index",
+                description="Integer index of the reference (0, 1, ...) that was predicted by the model (for "
+                "multiple-choice).",
+                lower_is_better=None,
+                group=None,
+            ),
+        ]
 
-    def derive_per_instance_stats(self, per_instance_stats: Dict[Instance, List[Stat]]) -> List[Stat]:
-        """Derive calibration metrics if applicable. We don't worry about splits and perturbations here."""
-        derived_stats: List[Stat] = []
-        derived_stats.extend(compute_calibration_metrics(per_instance_stats))
-        derived_stats.append(Stat(MetricName("num_instances")).add(len(per_instance_stats)))
-        return derived_stats
+
+def compute_request_state_metrics(
+    efficiency_metric: EfficiencyMetric,
+    adapter_spec: AdapterSpec,
+    request_state: RequestState,
+    metric_service: MetricService,
+) -> List[Stat]:
+    """
+    Compute metrics that are common to both `evaluate_generation` and `evaluate_references`.
+    """
+    stats: List[Stat] = []
+
+    stats.append(Stat(MetricName("num_references")).add(len(request_state.instance.references)))
+
+    # Copy from adapter spec
+    stats.append(Stat(MetricName("num_train_trials")).add(adapter_spec.num_train_trials))
+
+    stats.extend(efficiency_metric.compute_efficiency_metrics(adapter_spec, request_state, metric_service))
+    stats.extend(_compute_finish_reason_metrics(adapter_spec, request_state, metric_service))
+    stats.extend(_compute_truncation_metrics(adapter_spec, request_state, metric_service))
+
+    return stats
+
+
+def get_request_state_metrics_metadata(
+    efficiency_metric: EfficiencyMetric,
+) -> List[MetricMetadata]:
+    metric_metadata = [
+        MetricMetadata(
+            name="num_references",
+            display_name="# ref",
+            description="Number of references.",
+            lower_is_better=None,
+            group=None,
+        ),
+        MetricMetadata(
+            name="num_train_trials",
+            display_name="# trials",
+            description="Number of trials, where in each trial we choose an independent, random set of training "
+            "instances.",
+            lower_is_better=None,
+            group="general_information",
+        ),
+    ]
+    return (
+        metric_metadata
+        + efficiency_metric.get_metadata()
+        + _get_finish_reason_metrics_metadata()
+        + _get_truncation_metrics_metadata()
+    )
+
+
+def _compute_finish_reason_metrics(
+    adapter_spec: AdapterSpec, request_state: RequestState, metric_service: MetricService
+) -> List[Stat]:
+    """Record how often generation finished due to reaching token limit, stop token(s), or end of text"""
+    assert request_state.result is not None
+    sequence = request_state.result.completions[0]
+    valid_reasons = [
+        "length",
+        "stop",
+        "endoftext",
+        "unknown",
+    ]
+    if sequence.finish_reason is None or sequence.finish_reason["reason"] not in valid_reasons:
+        reason = "unknown"
+    else:
+        reason = sequence.finish_reason["reason"]
+    return [
+        Stat(MetricName(f"finish_reason_{valid_reason}")).add(int(reason == valid_reason))
+        for valid_reason in valid_reasons
+    ]
+
+
+def _get_finish_reason_metrics_metadata():
+    return [
+        MetricMetadata(
+            name="finish_reason_endoftext",
+            display_name="finish b/c endoftext",
+            description="Fraction of instances where the the output was terminated because the end of text token "
+            "was generated.",
+            lower_is_better=None,
+            group=None,
+        ),
+        MetricMetadata(
+            name="finish_reason_length",
+            display_name="finish b/c length",
+            description="Fraction of instances where the the output was terminated because of the max tokens limit.",
+            lower_is_better=None,
+            group=None,
+        ),
+        MetricMetadata(
+            name="finish_reason_stop",
+            display_name="finish b/c stop",
+            description="Fraction of instances where the the output was terminated because of the stop sequences.",
+            lower_is_better=None,
+            group=None,
+        ),
+        MetricMetadata(
+            name="finish_reason_unknown",
+            display_name="finish b/c unknown",
+            description="Fraction of instances where the the output was terminated for unknown reasons.",
+            lower_is_better=None,
+            group=None,
+        ),
+    ]
+
+
+def _compute_truncation_metrics(
+    adapter_spec: AdapterSpec, request_state: RequestState, metric_service: MetricService
+) -> List[Stat]:
+    """
+    Record the number of training instances used in the prompt and whether
+    even the prompt needed to be truncated (once we hit zero training instances).
+    """
+    return [
+        Stat(MetricName("num_train_instances")).add(request_state.num_train_instances),
+        Stat(MetricName("prompt_truncated")).add(request_state.prompt_truncated),
+    ]
+
+
+def _get_truncation_metrics_metadata() -> List[MetricMetadata]:
+    return [
+        MetricMetadata(
+            name="num_train_instances",
+            display_name="# train",
+            description="Number of training instances (e.g., in-context examples).",
+            lower_is_better=None,
+        ),
+        MetricMetadata(
+            name="prompt_truncated",
+            display_name="truncated",
+            description="Fraction of instances where the "
+            "prompt itself was truncated (implies "
+            "that there were no in-context "
+            "examples).",
+            lower_is_better=None,
+        ),
+    ]
+
+
+def compute_language_modeling_metrics(
+    adapter_spec: AdapterSpec, request_state: RequestState, metric_service: MetricService
+) -> List[Stat]:
+    """Compute the logprob and normalization factors for the first completion"""
+    assert request_state.result is not None
+    sequence = request_state.result.completions[0]
+
+    # Remove the empty tokens (typically generated by the AI21 tokenizer in the beginning of the text)
+    #
+    # Some more details about AI21 tokenizer: If the input prompt begins with a space, then
+    # the tokenizer inserts an empty token to the beginning.
+    # e.g. " burying him" -> ["▁"(0,0), "▁burying"(0,8), "▁him"(8,12)].
+    # TODO(#1522): Update this comment once solved.
+    # Since this empty token is introduced by our chunking approach, we need to remove it.
+    tokens: List[Token]
+    if request_state.num_conditioning_tokens > 0 and sequence.tokens[0].text == "":
+        tokens = sequence.tokens[1:]
+    else:
+        tokens = sequence.tokens
+    pred_tokens = tokens[request_state.num_conditioning_tokens :]
+    logprob, num_perplexity_tokens, num_bytes = (
+        sum(token.logprob for token in pred_tokens),
+        len(pred_tokens),
+        get_num_bytes(pred_tokens),
+    )
+
+    return [
+        Stat(MetricName("logprob")).add(logprob),
+        Stat(MetricName("num_perplexity_tokens")).add(num_perplexity_tokens),
+        Stat(MetricName("num_bytes")).add(num_bytes),
+    ]
+
+
+def _get_language_modeling_metrics_metadata() -> List[MetricMetadata]:
+    return [
+        MetricMetadata(
+            name="logprob",
+            display_name="Log probability",
+            short_display_name="Logprob",
+            description="Predicted output's average log probability (input's log prob for language modeling).",
+            lower_is_better=False,
+        ),
+        MetricMetadata(
+            name="num_perplexity_tokens",
+            display_name="# tokens",
+            description="Average number of tokens in the predicted output (for language modeling, the input too).",
+            lower_is_better=None,
+        ),
+        MetricMetadata(
+            name="num_bytes",
+            display_name="# bytes",
+            description="Average number of bytes in the predicted output (for language modeling, the input too).",
+            lower_is_better=None,
+        ),
+    ]
 
 
 def _has_non_zero_valued_logprobs(per_instance_stats: Dict[Instance, List[Stat]]) -> bool:
@@ -943,3 +636,80 @@ def compute_calibration_metrics(per_instance_stats: Dict[Instance, List[Stat]])
             stats.append(Stat(MetricName("platt_ece_1_bin")).add(platt_ece_1_bin))
 
     return stats
+
+
+def _get_calibration_metrics_metadata() -> List[MetricMetadata]:
+    return [
+        MetricMetadata(
+            name="ece_10_bin",
+            display_name="10-bin expected calibration error",
+            short_display_name="ECE (10-bin)",
+            description="The average difference between the model's confidence and accuracy, averaged across 10 "
+            "bins where each bin contains an equal number of points (only computed for classification "
+            "tasks). Warning - not reliable for small datasets (e.g., with < 300 examples) because "
+            "each bin will have very few examples.",
+            lower_is_better=True,
+            group="calibration",
+        ),
+        MetricMetadata(
+            name="ece_1_bin",
+            display_name="1-bin expected calibration error",
+            short_display_name="ECE (1-bin)",
+            description="The (absolute value) difference between the model's average confidence and accuracy "
+            "(only computed for classification tasks).",
+            lower_is_better=True,
+            group="calibration_detailed",
+        ),
+        MetricMetadata(
+            name="selective_acc@10",
+            display_name="Accuracy at 10% coverage",
+            short_display_name="Acc@10%",
+            description="The accuracy for the 10% of predictions that the model is most confident on (only "
+            "computed for classification tasks).",
+            lower_is_better=False,
+            group="calibration_detailed",
+        ),
+        MetricMetadata(
+            name="selective_cov_acc_area",
+            display_name="Selective coverage-accuracy area",
+            short_display_name="Selective Acc",
+            description="The area under the coverage-accuracy curve, a standard selective classification metric "
+            "(only computed for classification tasks).",
+            lower_is_better=False,
+            group="calibration_detailed",
+        ),
+        MetricMetadata(
+            name="platt_coef",
+            display_name="Platt Scaling Coefficient",
+            short_display_name="Platt Coef",
+            description="Coefficient of the Platt scaling classifier (can compare this across tasks).",
+            lower_is_better=False,
+            group="calibration_detailed",
+        ),
+        MetricMetadata(
+            name="platt_intercept",
+            display_name="Platt Scaling Intercept",
+            short_display_name="Platt Intercept",
+            description="Intercept of the Platt scaling classifier (can compare this across tasks).",
+            lower_is_better=False,
+            group="calibration_detailed",
+        ),
+        MetricMetadata(
+            name="platt_ece_10_bin",
+            display_name="10-bin Expected Calibration Error (after Platt scaling)",
+            short_display_name="Platt-scaled ECE (10-bin)",
+            description="10-bin ECE computed after applying Platt scaling to recalibrate the model's predicted "
+            "probabilities.",
+            lower_is_better=True,
+            group="calibration_detailed",
+        ),
+        MetricMetadata(
+            name="platt_ece_1_bin",
+            display_name="1-bin expected calibration error (after Platt scaling)",
+            short_display_name="Platt-scaled ECE (1-bin)",
+            description="1-bin ECE computed after applying Platt scaling to recalibrate the model's predicted "
+            "probabilities.",
+            lower_is_better=True,
+            group="calibration_detailed",
+        ),
+    ]